Title: DASH: Dual-Branch Score Distillation for Guidance-Calibrated Compact Diffusion Models

Abstract

While parameter compression in class-conditional diffusion models has been explored, a significant limitation in output-level distillation remains underexamined: the unconditional score branch is left unsupervised. This lack of supervision causes the classifier-free guidance gap to be underdetermined in the student model. Because this discrepancy is magnified at each denoising step, the training process often converges to degenerate solutions where both branches yield identical predictions. Consequently, guidance becomes ineffective even when the output-level training loss is low.

To address this, we present DASH, a novel dual-branch distillation framework that applies independent supervision to both score branches. By imposing distinct constraints on each branch, DASH uniquely specifies target outputs for every training sample. Additionally, an anchor term is employed to regularize conditional predictions, aligning them with ground-truth noise. The framework also incorporates TIRT Transfer, a mechanism that transfers the teacher’s converged per-timestep importance curriculum to the student as a frozen prior. This approach removes the necessity for the student to relearn the curriculum within the constraints of limited distillation budgets.

We evaluated our method on CIFAR-10 and CIFAR-100 datasets. Results show that with a 5.9x compression ratio, the model maintains quality within 4 FID points of the teacher during 50-step DDIM sampling. This performance significantly surpasses training from scratch, while effectively preserving guidance fidelity. Ablation studies highlight that unconditional supervision is the primary driver of improvement, contributing more than 60% to the total distillation gain. Together, curriculum transfer and anchor regularization offer complementary advantages, empirically validating that dual-branch constraints are essential for compression that preserves guidance.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC